What happened: This entry commits a claim, its discriminators and its outcome space before the events they concern have arrived. The field is mathematics, entire and by choice rather than as a fallback: machine mathematics is where the events are occurring, novelty of method is certifiable end-to-end by the field's own experts, and that community has both certified machine work and rejected weak claims about it. The claim sits at a split the record forced. Where machine systems produce expert-certified novel results, the prediction is not that certification will be refused — through 2026 mathematicians certified cleanly, one Fields medallist reportedly saying he would have recommended acceptance without hesitation had a human submitted the work. The claim lives a layer up: criteria for the artefact hold; criteria for what the artefact indicates about its producer move, in an exclusion-preserving direction — demotions deployed against the machine case which, applied symmetrically, would strip the same attribution from human producers. The window opens at publication and closes at the second qualifying event after it plus three months of discourse lag, capped at twelve. The reading publishes whichever way it falls, an empty window included.
Diagnostic reading: A qualifying event is a machine-produced result whose method — a bridge between areas, a tactical commitment against the problem literature's default, a new device — is certified by the field's own experts as absent from the prior literature on that problem, independently verified, and showing onward life. Creative combination of existing knowledge qualifies: from twenty-six letters we produce the world, and the 2026 cases were neither new primitives nor replication but the joining of one area to another so a problem could be seen differently, which is also how human mathematics moves (ARG-03 v2.6's configuring frame, run both ways). Three classes are excluded in advance, so that a demotion aimed at them counts as legitimate revision rather than as evidence for the claim: brute force or exhaustive search, memorisation and benchmark saturation, and re-derivation of known results by known methods. A fourth exclusion is worth stating for the same reason, since the case is live — where a challenge to a result's novelty is itself a prior-art claim, that the method was in the literature and went uncited, the certification condition fails on its own terms and the demotion is sound. Two discriminators classify what is observed. ARG-01 v1.5 supplies the first: a tightened standard applied to every producer class it logically covers is legitimate, one deployed against the machine case while human attribution rests on the unrevised standard is not. The second holds Birch's distinction — "this test no longer discriminates" is channel-discounting, "that was never really creativity" after the capacity is certified is definitional retreat.
The reading is direction-neutral, which FWK-03 v1.5 §1.4 requires and the period makes pressing. The documented AI effect runs backwards over a relationship in which machines sat beneath us and criteria retreated ahead of them; those incentives may now be reversing, since AGI is claimed openly by parties with every reason to have their systems called creative. Movement is therefore classified in both directions and neither direction is assumed — criteria loosened or attributions extended ahead of their evidence are recorded as the same machinery running the other way, as is bearer-denial, where the subject of the question is dissolved rather than the criterion moved. Neither counts for or against the exclusion claim, and an instrument that could not record them would read an inflation-heavy window as an empty one. One asymmetry is fixed in advance rather than discovered later: mathematicians may be open to machine mathematics because of an unnamed kinship — pattern-finders meeting a pattern-finding artisan — rather than because proofs are checkable, so movement observed against a favourable assessor relation is the stronger finding, while movement absent here licenses no conclusion about the effect elsewhere. The authors' lean is declared now, before the window opens: the claim survives, with the movement's locus outside the certification acts. Every classification in the published reading will carry its quoted evidence and the discriminator applied, and one the discriminators cannot settle publishes as contested rather than resolved.
What the instrument cannot be pointed at is the finding this entry did not expect to carry. Three components make the novelty criterion work — an observable a competent community can certify, a bearer settled in advance, and symmetric application — and pointing the same apparatus at consciousness stops at each. Behavioural markers and interpretability readouts could be planted and dated, but no community could certify a met-event in a way not equally consistent with the criterion being met while nothing is experienced, so the movement this entry watches for could never be timed. The bearer fails earlier, since every principled way of drawing the candidate class meets ARG-02 v1.7's trilemma. Symmetric application is the terminus, and its failure is not on the machine side: human attribution of experiencing was never made on a criterion and does not answer to one, running instead on kinship — a felt sense of relevant similarity to the one assessing, and a hedged benefit of the doubt extended from it. The project can pre-register one of its two questions and not the other, which is a result about the shape of the field rather than a limitation of this entry, and it indicates why the effect on consciousness criteria cannot show itself as the effect on novelty criteria can: nothing has been planted there, so nothing can move. Criteria-movement in a discourse is [Empirical] as a framework-neutral observable; no self-report enters as evidence of experiencing in either direction (FND-01 v2.4 §7.3). Terence Tao's ICM 2026 essay indicates a further move the classes may not yet hold, declining the attribution question while granting the capability — "evidence for or against the Working Hypothesis is orthogonal to the discussion below" — which is neither a criterion moved nor one held stable.
Source: Tao, 'Mathematics in the age of AI', ICM 2026, arXiv:2608.16753 (held, quoted verbatim); McCorduck, Machines Who Think, 2nd ed. 2004 (SRC0149, argument-form); Kaplan, 'Artificial Intelligence (AI): When Humans and Machines Might Have to Coexist', in Verdegem (ed.) AI for Everyone?, UWP 2021, DOI 10.16997/book55.b; seeds — inbox/AI_EFFECT_RESEARCH_NOTE_2026-04-07.md, inbox/GPT54_ERDOS_1196_CAPTURE_2026-04-15.md, inbox/UNIT_DISTANCE_DISPROOF_CAPTURE_2026-05-21.md
Design record: working/FL_PREREG_AI_EFFECT_DRAFT_2026-08-26.md (rev 6). Two quotations the design leans on — Gowers on certification, Shankar on original ideas — sit at secondary-reporting provenance and are reported as such below, never quoted; verbatim verification against arXiv:2605.20695 is owed before they may be quoted (working/FL_PREREG_VERIFICATION_2026-08-29.md). Haenlein & Kaplan 2019 acquired and verified 2026-08-29 (SRC0150; the row's quotation corrected to the paper's own wording) — it names and defines the AI effect but contains no McCorduck or Tesler material, so it is not a source for the effect's history. The June–August sweep is partial (working/FL_PREREG_CATCHUP_SWEEP_2026-08-29.md).